A bank reminder, a delivery update and a scammer impersonating a bank can now use the same basic technology: an automated voice agent capable of speaking, listening and responding in real time. As AI-generated calls become more common, the old distinction between a legitimate business call and a spam call is becoming harder to make.

The Telecom Regulatory Authority of India (Trai) is now preparing amendments to its Telecom Commercial Communications Customer Preference Regulations (TCCCPR) that would require businesses using automated calling systems to declare such calls to their telecom service providers. The amendments, first proposed for consultation in March, are now being finalised, according to an Economic Times report.

The timing is important. Telecom operators are already using AI-based systems to identify suspicious calling behaviour. In the April-June quarter of FY2026-27, these systems flagged 22.99 billion incoming calls as suspected spam, according to government data.

At the same time, businesses are beginning to use AI voice agents at considerable scale. Sarvam says its systems handle more than 2 million voice conversations a day, while Mahindra Finance and Sarvam recently said their voice agents had completed more than 10 million calls across 12 Indian languages.

That creates a new problem for India's anti-spam architecture: how does a network distinguish a machine that has been authorised to call from a machine that is being used to spam, scam or impersonate someone?

A call can be automated without being spam

The first thing to understand is that an automated call is not automatically an illegal or unwanted call.

A traditional robocall may play a recorded message, while newer AI voice agents can hold a conversation. A bank could use one to remind a customer about an upcoming repayment. An insurer could use one to discuss a renewal. A delivery company could call to confirm an address.

In each case, the call may have been initiated by software rather than a person.

Trai's proposed framework attempts to address this at the network level by bringing Application-to-Person, or A2P, calls into a more formal framework. The March consultation proposed that senders using A2P calls declare this usage to the originating telecom operator. Undeclared A2P traffic could be treated as unsolicited commercial communication.

The consultation is significant because A2P is broader than the old idea of a recorded robocall. The definition covers calls initiated through an application, software system or automated platform.

So how can a telecom operator know a call is legitimate?

There are several layers.

Identity

Trai has already been moving towards designated numbering resources for commercial voice calls. The 140 series is used for promotional calls, while 1600-series numbers have been allocated for service and transactional calls for specified sectors. In August, Trai also directed the allocation and operationalisation of the 1601 series for service and transactional calls by entities outside the BFSI and government sectors, including sectors such as utilities, courier and logistics.

This gives the network a basic signal: who is calling, under which numbering resource, and for what category of communication.

Registration and declaration

The amendments being finalised would require businesses using automated calling systems to declare that traffic to their telecom providers. That information can then be used by operators when deciding whether traffic from a particular sender is expected and authorised.

Behaviour

This is where AI already plays a major role. Telecom operators analyse calling patterns to identify numbers behaving like spam sources. A number making an unusually large volume of calls, repeatedly contacting people who have no apparent relationship with the sender, or showing other suspicious patterns can be flagged.

Trai said that such systems flagged 22.99 billion suspected spam calls during Q1 FY 2026-27.

Consumer feedback:

In 2025, consumers filed 3.1 million Unsolicited Commercial Communication (UCC) complaints across channels, including 1.7 million through the Trai DND app. Trai also issued more than 7.31 lakh notices to unregistered telemarketers during the year.

That feedback can become another signal for an operator or caller-ID platform.

Can AI tell whether the voice itself is AI-generated?

This is where things become considerably more complicated.

A telecom network does not need to determine whether a voice is synthetic in order to identify a spam caller. Network-level detection can work from metadata, number identity, call volumes, routing information, complaints and calling patterns.

That is useful because identifying the technology behind the voice is not necessarily the same as identifying the risk.

A scammer could use an ordinary human caller with a spoofed number. Another could use a perfectly natural-sounding AI voice. Both can be fraudulent.

Caller-ID platforms and smartphones are beginning to add another layer by analysing what happens during the call itself.

Truecaller, for example, now offers an AI Call Scanner that it says can detect whether a caller is a human or an AI-synthesised voice by analysing a voice sample in real time.

Google has also introduced on-device Scam Detection feature for select Pixel smartphones. On Pixel 9 and later devices in India, its on-device AI model (Gemini Nano) can analyse calls for scam indicators while the conversation is taking place. Google cautions that the system is not 100 per cent accurate because scammers continuously change their tactics.

This points to an important distinction. Detecting an AI voice and detecting a scam are two different problems.

An AI voice can belong to a legitimate bank. A human voice can belong to a scammer. A scammer can also combine a real person's voice clone with a spoofed caller ID.

Why caller ID alone cannot solve the problem

Caller ID remains useful, but it is not proof of authenticity. Fraudsters can spoof numbers or make a call appear to originate from a trusted organisation. This is why regulators in other markets have increasingly focused on verifying the underlying origin of calls rather than relying only on the number displayed to consumers.

The UK telecom regulator Ofcom, for example, has strengthened requirements around calls from abroad that display spoofed UK numbers. Providers are expected to block certain calls where the displayed number cannot be verified as legitimate.

India's designated numbering approach follows a related principle: make legitimate commercial traffic easier to identify at the network level.

For consumers, that creates a useful rule of thumb. A designated number or verified caller identity can increase confidence, but it should not become a substitute for checking what the caller is asking for.

Where does India stand globally?

India's approach is currently centred heavily on the telecom network. Trai is building a system around registered senders, designated number series, consent, DND preferences, complaint mechanisms and AI-based detection. Its February 2026 direction also institutionalised AI/ML-based UCC_Detect intelligence for sharing suspected spam information between operators.

The US has taken a somewhat different route. In 2024, the Federal Communications Commission ruled that AI-generated voices fall within the definition of "artificial or prerecorded voice" under the Telephone Consumer Protection Act. This brought AI voice-cloning robocalls within existing restrictions, including consent requirements.

The European Union has gone further on disclosure. Article 50 of the EU AI Act requires providers of AI systems intended to interact directly with people to ensure that people are informed that they are interacting with an AI system, unless the interaction is obvious from the circumstances. The transparency obligations started applying on August 2, 2026.